# Mechanica

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Jim Huang](https://cerebralvalley.ai/u/jimtw)
- **GitHub:** https://github.com/jhuang-tw/mechanica
- **Demo video:** https://youtu.be/Ata6pzX-ook
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1543

Mechanica is an interactive machine classroom where Meta Muse Spark is the teacher: it repairs, X-rays and takes apart real machines in 3D, on your phone or in a Quest headset. Live demo: https://mechanica.hoson.xyz

The problem. People learn machines with their hands: open the engine and watch the gears turn. Most learners never get that chance. Real machines are expensive, dangerous or sealed, textbooks show static cutaways, and videos can't answer "why".

What Mechanica does. Eight machines modelled from scratch during the hackathon (road bicycle, car, steam locomotive, helicopter, turbofan jet, mechanical dive watch, mixed-reality headset, 700 cc motorbike; 282 moving parts) become a classroom where Meta Muse Spark is the teacher:

1. Repair challenge (home screen). A broken machine runs in front of you. Say, type or tap what you think is wrong. Muse Spark checks your guess on the 3D model with tool calls (isolate the part, run the real drive, turn the shell to X-ray glass) and gives a clue, not the answer. When you find it, you repair it, explain why in your own words, and Muse grades the explanation and replays only the cause-and-effect link you missed. Every tool call appears on screen as it plays.
2. Muse tour. One tap and Muse Spark directs a 30–60 second narrated teardown, calling a scene tool 3–8 times for camera moves, explode, X-ray, isolated parts and running drives.
3. X-ray teardown. Taking a machine apart turns its shell to glass, so the mechanism stays in view while it runs.
4. Photo to 3D. Photograph a real machine. Muse Spark vision names it and its visible parts, SAM 3.1 traces each part on the photo, and a tapped trace flies out of the photo into the matching part of the exploded 3D model.
5. Ask Muse, lessons and tutor. Ask anything and the answer is a demonstration Muse Spark choreographs with the exhibit's own controls. 25 cause-and-effect lessons ask you to guess before each step; a Muse tutor grades free spoken answers and remembers weak spots.
6. Hands-on and VR. Drag a pedal or piston and the whole drive follows at real gear ratios. The same URL runs in WebXR on Meta Quest: wrist menu, real hand meshes, two fists to explode and turn, voice questions through Muse Voice Transcribe.

Who it is for. Vocational students, junior technicians, STEM classrooms and anyone curious, with nothing to install. English and Traditional Chinese. A typical session: a student opens the bicycle on a phone, sees the pedals spin while the wheel stays still, guesses "the brake", and Muse runs the drive with the chain isolated to show the force never reaches the rear sprocket. The student finds the slipped chain, puts it back, and explains in one sentence why the wheel now turns. A school without a bike, a jet engine or a steam locomotive on the bench gets the same fault-finding practice, and a teacher can project it or put the class in headsets.

Under the hood. Each exhibit is one JSON manifest (parts, real drives with gear ratios, linkages, lessons, faults) that is also the schema for Muse's tools, so the model cannot name a part or move a value the machine doesn't have. The eight models are generated by scripts (build123d CAD for gears, cranks and casings, Blender for assembly and shells, a GLB optimizer for phones), one node per manifest part. A repair turn is two Muse calls with tool runs in between and takes about 4–6 seconds. 39 offline tests cover the rig, gestures, the photo matching, the tour validator, the server and both languages.

Built during the hackathon. The public repository was created on 2026-10-27 and every commit, model and asset in it dates from the hackathon window. docs/BUILT_DURING_HACKATHON.md lists each feature with the Meta model it uses, its commits and its date.

## More from Meta Global AI Developer Hackathon

- [FormerLamp](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1540)
- [1124607440057522](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1541)
- [Fluixi AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1542)
- [saket kunwar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1544)
- [AuraMatrix AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1545)
- [Coco_vSp](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1546)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1543. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt